Generalized Multilevel Model with multinominal distribution



Hi all,

I have multilevel data with a categorical dependent variable and doubt about which Link function I should select in SPSS with the Multinomial Distribution Function (i.e. complementary log-log, cuachit, logit, negative log-log, probit). The independent variable is not normally distributed.

The data are about the confusions (mistakes) participants make in an emotion recognition task. So, they might say the stimulus picture they looked at showed anger, but in fact it showed fear. Participants saw stimuli from four emotion categories and could also choose among four response options. As I only want to analyze the mistakes, I have the following data (Emotion Category (independent variable) - Emotion Response (dep. var)):
(anger-sad) (anger-anger is thus excluded, making the iv not normal!)

In other words, the Emotion Category includes different levels per response category. In case the response was incorrect and was 'angry', the emotion category could have been anything but angry and thus only includes stimuli showing happy, sad and fearful expressions. But for the fear responses, the Emotion Category variable inluces the emotions anger, happy and sad (and not fear). So there is always one level missing.

Thanks for your advice!

best wishes,


You may use random intercepts and random slopes when this data come form the same obserwations. Also you can compare models with random slopes and fixed slopes by AIC coeffitient. Lower AIC indicate better predictive model :)